The Human Brain Project

[pygmy_giant]

http://www.humanbrainproject.eu/future_computing.html

Fascinating.

[morphy_richards]

Need to read this in more detail but , from a quick scan the description of the sheer computing power necessary for the project, I would imagine the American military and others would have an issue because that could / would be classed as a weapon. Additionally it sounds like it would be powerful enough to crack long encryption keys in unacceptably short amounts of time.
Think I had better shut up and read it properly before I make myself look like a numpty (again!)

[shawaj]

interesting project.

@morphy - agreed with your comments, but also think the US military need to stop thinking they are the world police and can control everyone else!

[morphy_richards]

Another impressive aspect, aside from the sheer geekattraction... If you look under "European partners", the sheer number of jobs that have been created in an area that not long ago only existed in Ian M Banks culture novels.

[pygmy_giant]

On the down side the machines are taking over and could pose a threat to all our jobs - I say we target the headquaters of this project with a R.O.N. (Reality or Nothing) terrorist strike.

http://www.flickr.com/photos/pkdouyk/244835186/

(Joking of course) Err.. is that insightment to violence? I didn't mean it - don't prosecute me under the terroism act - please - it was a joke - honest - look - I'm smiling - you can see I'm smiling - can't you? There, see? Its was a sarcastic ironic smile - you can see that can't you? I'm being ironic - I don't condone violence - and thats thats the truth - not irony see - serious face.

Phew! Think I got away with that obviously ironic not insighting violence joke. Ha. Ha. Life is so funny.

Seriously though, I was wondering about how to simulata a brain on the Pi.

I think you could do it with a massive square shaped 2D array representing a large number of neurons and all the possible connections between them.

In this array you would keep the synapse weighting of the interconnections ranging from 0 (no connection) to a maximum value.

There would be 2 additional 1 dimensional buffer arrays each as long as the side of the square connection array whose length relates to the number of neurons. These could contain the summed input of the outputs of the firing connected neurons.

There would also be a second similar 1D array representing the firing threshold of each neuron.

I envisage each cycle of the simulation program checking the threshold of each neuron against input buffer 1 to ascertain whether this input strength is sufficient to switch it on into firing mode and then following the associated connections in the 2D array to add the firing signal strength to the buffer array 2.

After this is done, input buffer 1 would be transferred to buffer 2 (or use assignment flags swapped) before input buffer 1 is cleare and the whole process is repeated.

The fun part would be assigning initial stae conditions and tweaking those according to a training program in relation to a simple pre-chosen task and working out how to input and output from the system.

[morphy_richards]

I've had a kind of similar idea bobbing around in the back of my mind for a while, except my version was going to use self referencing tables in a mySQL database... As it's file based you could hook a raspi to a really big external hard disk and simulate quite a large network. As mySQL is good at optimising transactions on that sort of thing I thought it might be the best way forwards
Except I cant remember much about neural networks anymore so it's all just sort of a vague notion currently .

Robot wise, there might be some shortcuts where you can dispense with the neurons and use some other kind of learning algorithms but having said that nothing has happened with my pet "skutter" for rather a long time now. Stupid real life and real people, why wont they all just leave me alone and let me get on with my tinkering?

[riffraff]

Just a few thoughts I've been tossing around lately, y'all decide whether or not it's BS. Everyone may be making the problem too difficult by trying to emulate wetware functions in hardware. The key may be playing to the strengths of hardware rather than forcing it into such an awkward model. We need to be looking at the functions of the brain at the block diagram instead of the schematic level: Sensory Processing, Knowledge Base, Knowledge Index, Pattern Association Engine, Behavioral Schema, Command Processing, Procedural Analysis and Compilation. Break the problem down to the basics and begin to look at how it can be implemented in networked system. We're already accomplishing some of this this on a Meta-system level with the web, the symbology involved is just too complex and needs scaling down to a binary level to approach speeds necessary for useful "cognition".

Sensory Processing:
We've already got a head start here with pattern recognition systems, facial recognition, etc. Google's gone a long way here with systems that can recognize the contents of a photo. We've had voiceprint and sound recognition for years. That's key in modern sonar listening systems and apps such as Sound Hound. Smell sensors have recently been designed that can distinguish the unique odors produced by infectious organisms and are currently being used in medical diagnosis.

Knowledge Base:
According to Global Language Monitor there are currently 1,013,913 words in the English language. A 32 bit system can uniquely allocate up to 4,294,967,296 individual address to storage nodes - that's a fabulous start. When we think of how information is stored in the human brain, it may not be immediately necessary to cross-connect every storage node to achieve the desired effect in hardware. it may be more effective to store basic definitions for objects and verbs in the most primitive terms along with associated sensory matching templates, procedures and detailed images and descriptions. "Concepts" involve linkage of multiple primitives such as "throw" and "ball" are conjoined to produce "throwing a ball" and obtain their own allocation with associated images and procedures.

Knowledge Index:
The equivalent to the human hippocampus. Primarily, an allocation table that lists the locations of primitives and concepts. Secondarily, a repository for higher level functions involving the linkage of objects and procedures resulting from posited or learned behavior, i.e. "throwing a ball to first base to get the out". This is where your numbers get really big. The larger the index value, the more complex the system. 32 bits ain't enough, you're looking at 128, 256, 1024 or higher.

Pattern Association Engine:
This is the basic equivalent to a "Web-Crawler", and this is the cutting edge of algorithmic development where things get hairy. This would be like human free-association combined with the mysterious problem solving mechanism that works during REM sleep. This is where you want to identify duplicated or similar sets of matching templates or procedures that represent analogous concepts. At the very least, this reduces the complexity of the Procedural Analysis process, but with time and perfection, could yield oblique reasoning and independent discovery.

Behavioral Schema:
You'd definitely put the three laws here, along with any other limiting functions to prevent potential disaster.

Command Processing:
Parsing of immediate instructions along with preset or compiled procedures required for regular service or self sustenance

Procedural Analysis and Compilation:
The back-end of Command Processing where, once identified, the "brain" begins to search the Knowledge Index for object definitions and procedures that will fulfill the assigned commands and place them into a coherent executable instruction stream.

[riffraff]

This is cool. Check it out :
http://www.dailytech.com/Googles+Unsupe ... e25025.htm

Seems like it had no trouble recognizing cat anatomy, after all, a cat is a cat, is a cat. A little more problematic with human faces, however. I have to wonder, though, what would have happened if they had incorporated the mathematical facial recognition algorithms developed for security. Would it then have been able to differentiate individual faces, or males from females, etc.?

They have what, 16K processors interconnected? If the process is scaled down to hardware level, that can certainly reduce the system requirements, and I think you'd all agree that an SOC and memory combo that occupies less than half a square inch of board real estate, like with the Pi, that the hardware size and power requirements can be vastly reduced.

[morphy_richards]

Did anyone ever come across the Gleisner robots in Greg Egan novels?
Basically, you have a blank artificial brain "attached" to you at some early point in your development. It is bootstrapped to all of your senses and copies your real brain's responses. It experiences everything that you experience ... as time goes on this artifical brain learns everything your actual brain does and becomes a perfect copy of your mind. When you reach an age where your real brain starts to degrade, it is hoiked out and the artificial brain takes over completely.

Is it still you or do you get murdered and replaced like in 'inavasion of the body snatchers'?

How about this scenario then: You have a degenerative brain disease. One small part of your brain becomes damaged (say the Basal Ganglia), so this is removed and replaced with an artificial part. Are you still you? (I personally would say you are)

Now lets say some time later the disease spreads to a neighbouring part of the brain - this is then removed and replaced with an artificial part. Are you still you now?

You can probably see where this is going ... the disease keeps spreading and parts keep getting replaced ... eventually the entire biological brain is replaced with artificial brain. At which point, if any, did you stop being you and become a "robot body snatcher"?

Anyway, getting back onto the point. These Gleisner robots don't seem so much like unrealistic SF when taking account of this "human brain project" and when coupled with developments in things like augmented relaity (google glasses etc.) it doesnt seem quite so far fetched to me.

[Ravenous]

It is bootstrapped to all of your senses and copies your real brain's responses.

That means it would need to detect not only all of your sensory inputs, but every connection within your brain. All of the internal state as well as the ins & outs.

How would they fit in all the wires?

As for changing after the brain is modified - well everything that happens to your brain (and possibly the rest of your hardware) affects you and the way you think, irreversibly. New brain connections are formed all the time apparently. Many things I found funny as a child, I'd be ashamed to admit these days. I really can't believe the sort of person I was...

[morphy_richards]

New brain connections are formed all the time apparently.

It's been a while, I don't get to read for pleasure much these days but I think that was covered too. It learned to form new connections in exactly the same way as your real brain did / does.

Anyway, assume you had some sort of technology (maybe not based on silicon) that could duplicate or exceed the scale and number of connections in a human brain (which could be an off-shoot of this project one day in the future) ...

Many things I found funny as a child, I'd be ashamed to admit these days

I'm ashamed to admit that these days many things I found funny as a child I still find funny .